cfl.c (16198B)
1 /* 2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved. 3 * 4 * This source code is subject to the terms of the BSD 2 Clause License and 5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 6 * was not distributed with this source code in the LICENSE file, you can 7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 8 * Media Patent License 1.0 was not distributed with this source code in the 9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 10 */ 11 12 #include "av1/common/av1_common_int.h" 13 #include "av1/common/cfl.h" 14 #include "av1/common/common_data.h" 15 16 #include "config/av1_rtcd.h" 17 18 void cfl_init(CFL_CTX *cfl, const SequenceHeader *seq_params) { 19 assert(block_size_wide[CFL_MAX_BLOCK_SIZE] == CFL_BUF_LINE); 20 assert(block_size_high[CFL_MAX_BLOCK_SIZE] == CFL_BUF_LINE); 21 22 memset(&cfl->recon_buf_q3, 0, sizeof(cfl->recon_buf_q3)); 23 memset(&cfl->ac_buf_q3, 0, sizeof(cfl->ac_buf_q3)); 24 cfl->subsampling_x = seq_params->subsampling_x; 25 cfl->subsampling_y = seq_params->subsampling_y; 26 cfl->are_parameters_computed = 0; 27 cfl->store_y = 0; 28 // The DC_PRED cache is disabled by default and is only enabled in 29 // cfl_rd_pick_alpha 30 clear_cfl_dc_pred_cache_flags(cfl); 31 } 32 33 void cfl_store_dc_pred(MACROBLOCKD *const xd, const uint8_t *input, 34 CFL_PRED_TYPE pred_plane, int width) { 35 assert(pred_plane < CFL_PRED_PLANES); 36 assert(width <= CFL_BUF_LINE); 37 38 if (is_cur_buf_hbd(xd)) { 39 uint16_t *const input_16 = CONVERT_TO_SHORTPTR(input); 40 memcpy(xd->cfl.dc_pred_cache[pred_plane], input_16, width << 1); 41 return; 42 } 43 44 memcpy(xd->cfl.dc_pred_cache[pred_plane], input, width); 45 } 46 47 static void cfl_load_dc_pred_lbd(const int16_t *dc_pred_cache, uint8_t *dst, 48 int dst_stride, int width, int height) { 49 for (int j = 0; j < height; j++) { 50 memcpy(dst, dc_pred_cache, width); 51 dst += dst_stride; 52 } 53 } 54 55 static void cfl_load_dc_pred_hbd(const int16_t *dc_pred_cache, uint16_t *dst, 56 int dst_stride, int width, int height) { 57 const size_t num_bytes = width << 1; 58 for (int j = 0; j < height; j++) { 59 memcpy(dst, dc_pred_cache, num_bytes); 60 dst += dst_stride; 61 } 62 } 63 void cfl_load_dc_pred(MACROBLOCKD *const xd, uint8_t *dst, int dst_stride, 64 TX_SIZE tx_size, CFL_PRED_TYPE pred_plane) { 65 const int width = tx_size_wide[tx_size]; 66 const int height = tx_size_high[tx_size]; 67 assert(pred_plane < CFL_PRED_PLANES); 68 assert(width <= CFL_BUF_LINE); 69 assert(height <= CFL_BUF_LINE); 70 if (is_cur_buf_hbd(xd)) { 71 uint16_t *dst_16 = CONVERT_TO_SHORTPTR(dst); 72 cfl_load_dc_pred_hbd(xd->cfl.dc_pred_cache[pred_plane], dst_16, dst_stride, 73 width, height); 74 return; 75 } 76 cfl_load_dc_pred_lbd(xd->cfl.dc_pred_cache[pred_plane], dst, dst_stride, 77 width, height); 78 } 79 80 // Due to frame boundary issues, it is possible that the total area covered by 81 // chroma exceeds that of luma. When this happens, we fill the missing pixels by 82 // repeating the last columns and/or rows. 83 static inline void cfl_pad(CFL_CTX *cfl, int width, int height) { 84 const int diff_width = width - cfl->buf_width; 85 const int diff_height = height - cfl->buf_height; 86 87 if (diff_width > 0) { 88 const int min_height = height - diff_height; 89 uint16_t *recon_buf_q3 = cfl->recon_buf_q3 + (width - diff_width); 90 for (int j = 0; j < min_height; j++) { 91 const uint16_t last_pixel = recon_buf_q3[-1]; 92 assert(recon_buf_q3 + diff_width <= cfl->recon_buf_q3 + CFL_BUF_SQUARE); 93 for (int i = 0; i < diff_width; i++) { 94 recon_buf_q3[i] = last_pixel; 95 } 96 recon_buf_q3 += CFL_BUF_LINE; 97 } 98 cfl->buf_width = width; 99 } 100 if (diff_height > 0) { 101 uint16_t *recon_buf_q3 = 102 cfl->recon_buf_q3 + ((height - diff_height) * CFL_BUF_LINE); 103 for (int j = 0; j < diff_height; j++) { 104 const uint16_t *last_row_q3 = recon_buf_q3 - CFL_BUF_LINE; 105 assert(recon_buf_q3 + width <= cfl->recon_buf_q3 + CFL_BUF_SQUARE); 106 for (int i = 0; i < width; i++) { 107 recon_buf_q3[i] = last_row_q3[i]; 108 } 109 recon_buf_q3 += CFL_BUF_LINE; 110 } 111 cfl->buf_height = height; 112 } 113 } 114 115 static void subtract_average_c(const uint16_t *src, int16_t *dst, int width, 116 int height, int round_offset, int num_pel_log2) { 117 int sum = round_offset; 118 const uint16_t *recon = src; 119 for (int j = 0; j < height; j++) { 120 for (int i = 0; i < width; i++) { 121 sum += recon[i]; 122 } 123 recon += CFL_BUF_LINE; 124 } 125 const int avg = sum >> num_pel_log2; 126 for (int j = 0; j < height; j++) { 127 for (int i = 0; i < width; i++) { 128 dst[i] = src[i] - avg; 129 } 130 src += CFL_BUF_LINE; 131 dst += CFL_BUF_LINE; 132 } 133 } 134 135 CFL_SUB_AVG_FN(c) 136 137 static inline int cfl_idx_to_alpha(uint8_t alpha_idx, int8_t joint_sign, 138 CFL_PRED_TYPE pred_type) { 139 const int alpha_sign = (pred_type == CFL_PRED_U) ? CFL_SIGN_U(joint_sign) 140 : CFL_SIGN_V(joint_sign); 141 if (alpha_sign == CFL_SIGN_ZERO) return 0; 142 const int abs_alpha_q3 = 143 (pred_type == CFL_PRED_U) ? CFL_IDX_U(alpha_idx) : CFL_IDX_V(alpha_idx); 144 return (alpha_sign == CFL_SIGN_POS) ? abs_alpha_q3 + 1 : -abs_alpha_q3 - 1; 145 } 146 147 static inline void cfl_predict_lbd_c(const int16_t *ac_buf_q3, uint8_t *dst, 148 int dst_stride, int alpha_q3, int width, 149 int height) { 150 for (int j = 0; j < height; j++) { 151 for (int i = 0; i < width; i++) { 152 dst[i] = clip_pixel(get_scaled_luma_q0(alpha_q3, ac_buf_q3[i]) + dst[i]); 153 } 154 dst += dst_stride; 155 ac_buf_q3 += CFL_BUF_LINE; 156 } 157 } 158 159 CFL_PREDICT_FN(c, lbd) 160 161 #if CONFIG_AV1_HIGHBITDEPTH 162 static inline void cfl_predict_hbd_c(const int16_t *ac_buf_q3, uint16_t *dst, 163 int dst_stride, int alpha_q3, 164 int bit_depth, int width, int height) { 165 for (int j = 0; j < height; j++) { 166 for (int i = 0; i < width; i++) { 167 dst[i] = clip_pixel_highbd( 168 get_scaled_luma_q0(alpha_q3, ac_buf_q3[i]) + dst[i], bit_depth); 169 } 170 dst += dst_stride; 171 ac_buf_q3 += CFL_BUF_LINE; 172 } 173 } 174 175 CFL_PREDICT_FN(c, hbd) 176 #endif 177 178 static void cfl_compute_parameters(MACROBLOCKD *const xd, TX_SIZE tx_size) { 179 CFL_CTX *const cfl = &xd->cfl; 180 // Do not call cfl_compute_parameters multiple time on the same values. 181 assert(cfl->are_parameters_computed == 0); 182 183 cfl_pad(cfl, tx_size_wide[tx_size], tx_size_high[tx_size]); 184 cfl_get_subtract_average_fn(tx_size)(cfl->recon_buf_q3, cfl->ac_buf_q3); 185 cfl->are_parameters_computed = 1; 186 } 187 188 void av1_cfl_predict_block(MACROBLOCKD *const xd, uint8_t *dst, int dst_stride, 189 TX_SIZE tx_size, int plane) { 190 CFL_CTX *const cfl = &xd->cfl; 191 MB_MODE_INFO *mbmi = xd->mi[0]; 192 assert(is_cfl_allowed(xd)); 193 194 if (!cfl->are_parameters_computed) cfl_compute_parameters(xd, tx_size); 195 196 const int alpha_q3 = 197 cfl_idx_to_alpha(mbmi->cfl_alpha_idx, mbmi->cfl_alpha_signs, plane - 1); 198 assert((tx_size_high[tx_size] - 1) * CFL_BUF_LINE + tx_size_wide[tx_size] <= 199 CFL_BUF_SQUARE); 200 #if CONFIG_AV1_HIGHBITDEPTH 201 if (is_cur_buf_hbd(xd)) { 202 uint16_t *dst_16 = CONVERT_TO_SHORTPTR(dst); 203 cfl_get_predict_hbd_fn(tx_size)(cfl->ac_buf_q3, dst_16, dst_stride, 204 alpha_q3, xd->bd); 205 return; 206 } 207 #endif 208 cfl_get_predict_lbd_fn(tx_size)(cfl->ac_buf_q3, dst, dst_stride, alpha_q3); 209 } 210 211 static void cfl_luma_subsampling_420_lbd_c(const uint8_t *input, 212 int input_stride, 213 uint16_t *output_q3, int width, 214 int height) { 215 for (int j = 0; j < height; j += 2) { 216 for (int i = 0; i < width; i += 2) { 217 const int bot = i + input_stride; 218 output_q3[i >> 1] = 219 (input[i] + input[i + 1] + input[bot] + input[bot + 1]) << 1; 220 } 221 input += input_stride << 1; 222 output_q3 += CFL_BUF_LINE; 223 } 224 } 225 226 static void cfl_luma_subsampling_422_lbd_c(const uint8_t *input, 227 int input_stride, 228 uint16_t *output_q3, int width, 229 int height) { 230 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 231 for (int j = 0; j < height; j++) { 232 for (int i = 0; i < width; i += 2) { 233 output_q3[i >> 1] = (input[i] + input[i + 1]) << 2; 234 } 235 input += input_stride; 236 output_q3 += CFL_BUF_LINE; 237 } 238 } 239 240 static void cfl_luma_subsampling_444_lbd_c(const uint8_t *input, 241 int input_stride, 242 uint16_t *output_q3, int width, 243 int height) { 244 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 245 for (int j = 0; j < height; j++) { 246 for (int i = 0; i < width; i++) { 247 output_q3[i] = input[i] << 3; 248 } 249 input += input_stride; 250 output_q3 += CFL_BUF_LINE; 251 } 252 } 253 254 #if CONFIG_AV1_HIGHBITDEPTH 255 static void cfl_luma_subsampling_420_hbd_c(const uint16_t *input, 256 int input_stride, 257 uint16_t *output_q3, int width, 258 int height) { 259 for (int j = 0; j < height; j += 2) { 260 for (int i = 0; i < width; i += 2) { 261 const int bot = i + input_stride; 262 output_q3[i >> 1] = 263 (input[i] + input[i + 1] + input[bot] + input[bot + 1]) << 1; 264 } 265 input += input_stride << 1; 266 output_q3 += CFL_BUF_LINE; 267 } 268 } 269 270 static void cfl_luma_subsampling_422_hbd_c(const uint16_t *input, 271 int input_stride, 272 uint16_t *output_q3, int width, 273 int height) { 274 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 275 for (int j = 0; j < height; j++) { 276 for (int i = 0; i < width; i += 2) { 277 output_q3[i >> 1] = (input[i] + input[i + 1]) << 2; 278 } 279 input += input_stride; 280 output_q3 += CFL_BUF_LINE; 281 } 282 } 283 284 static void cfl_luma_subsampling_444_hbd_c(const uint16_t *input, 285 int input_stride, 286 uint16_t *output_q3, int width, 287 int height) { 288 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE); 289 for (int j = 0; j < height; j++) { 290 for (int i = 0; i < width; i++) { 291 output_q3[i] = input[i] << 3; 292 } 293 input += input_stride; 294 output_q3 += CFL_BUF_LINE; 295 } 296 } 297 #endif 298 299 CFL_GET_SUBSAMPLE_FUNCTION(c) 300 301 #if CONFIG_AV1_HIGHBITDEPTH 302 static inline cfl_subsample_hbd_fn cfl_subsampling_hbd(TX_SIZE tx_size, 303 int sub_x, int sub_y) { 304 if (sub_x == 1) { 305 if (sub_y == 1) { 306 return cfl_get_luma_subsampling_420_hbd(tx_size); 307 } 308 return cfl_get_luma_subsampling_422_hbd(tx_size); 309 } 310 return cfl_get_luma_subsampling_444_hbd(tx_size); 311 } 312 #endif 313 314 static inline cfl_subsample_lbd_fn cfl_subsampling_lbd(TX_SIZE tx_size, 315 int sub_x, int sub_y) { 316 if (sub_x == 1) { 317 if (sub_y == 1) { 318 return cfl_get_luma_subsampling_420_lbd(tx_size); 319 } 320 return cfl_get_luma_subsampling_422_lbd(tx_size); 321 } 322 return cfl_get_luma_subsampling_444_lbd(tx_size); 323 } 324 325 static void cfl_store(CFL_CTX *cfl, const uint8_t *input, int input_stride, 326 int row, int col, TX_SIZE tx_size, int use_hbd) { 327 const int width = tx_size_wide[tx_size]; 328 const int height = tx_size_high[tx_size]; 329 const int tx_off_log2 = MI_SIZE_LOG2; 330 const int sub_x = cfl->subsampling_x; 331 const int sub_y = cfl->subsampling_y; 332 const int store_row = row << (tx_off_log2 - sub_y); 333 const int store_col = col << (tx_off_log2 - sub_x); 334 const int store_height = height >> sub_y; 335 const int store_width = width >> sub_x; 336 337 // Invalidate current parameters 338 cfl->are_parameters_computed = 0; 339 340 // Store the surface of the pixel buffer that was written to, this way we 341 // can manage chroma overrun (e.g. when the chroma surfaces goes beyond the 342 // frame boundary) 343 if (col == 0 && row == 0) { 344 cfl->buf_width = store_width; 345 cfl->buf_height = store_height; 346 } else { 347 cfl->buf_width = OD_MAXI(store_col + store_width, cfl->buf_width); 348 cfl->buf_height = OD_MAXI(store_row + store_height, cfl->buf_height); 349 } 350 351 // Check that we will remain inside the pixel buffer. 352 assert(store_row + store_height <= CFL_BUF_LINE); 353 assert(store_col + store_width <= CFL_BUF_LINE); 354 355 // Store the input into the CfL pixel buffer 356 uint16_t *recon_buf_q3 = 357 cfl->recon_buf_q3 + (store_row * CFL_BUF_LINE + store_col); 358 #if CONFIG_AV1_HIGHBITDEPTH 359 if (use_hbd) { 360 cfl_subsampling_hbd(tx_size, sub_x, sub_y)(CONVERT_TO_SHORTPTR(input), 361 input_stride, recon_buf_q3); 362 } else { 363 cfl_subsampling_lbd(tx_size, sub_x, sub_y)(input, input_stride, 364 recon_buf_q3); 365 } 366 #else 367 (void)use_hbd; 368 cfl_subsampling_lbd(tx_size, sub_x, sub_y)(input, input_stride, recon_buf_q3); 369 #endif 370 } 371 372 // Adjust the row and column of blocks smaller than 8X8, as chroma-referenced 373 // and non-chroma-referenced blocks are stored together in the CfL buffer. 374 static inline void sub8x8_adjust_offset(const CFL_CTX *cfl, int mi_row, 375 int mi_col, int *row_out, 376 int *col_out) { 377 // Increment row index for bottom: 8x4, 16x4 or both bottom 4x4s. 378 if ((mi_row & 0x01) && cfl->subsampling_y) { 379 assert(*row_out == 0); 380 (*row_out)++; 381 } 382 383 // Increment col index for right: 4x8, 4x16 or both right 4x4s. 384 if ((mi_col & 0x01) && cfl->subsampling_x) { 385 assert(*col_out == 0); 386 (*col_out)++; 387 } 388 } 389 390 void cfl_store_tx(MACROBLOCKD *const xd, int row, int col, TX_SIZE tx_size, 391 BLOCK_SIZE bsize) { 392 CFL_CTX *const cfl = &xd->cfl; 393 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y]; 394 uint8_t *dst = &pd->dst.buf[(row * pd->dst.stride + col) << MI_SIZE_LOG2]; 395 396 if (block_size_high[bsize] == 4 || block_size_wide[bsize] == 4) { 397 // Only dimensions of size 4 can have an odd offset. 398 assert(!((col & 1) && tx_size_wide[tx_size] != 4)); 399 assert(!((row & 1) && tx_size_high[tx_size] != 4)); 400 sub8x8_adjust_offset(cfl, xd->mi_row, xd->mi_col, &row, &col); 401 } 402 cfl_store(cfl, dst, pd->dst.stride, row, col, tx_size, is_cur_buf_hbd(xd)); 403 } 404 405 static inline int max_intra_block_width(const MACROBLOCKD *xd, 406 BLOCK_SIZE plane_bsize, int plane, 407 TX_SIZE tx_size) { 408 const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane) 409 << MI_SIZE_LOG2; 410 return ALIGN_POWER_OF_TWO(max_blocks_wide, tx_size_wide_log2[tx_size]); 411 } 412 413 static inline int max_intra_block_height(const MACROBLOCKD *xd, 414 BLOCK_SIZE plane_bsize, int plane, 415 TX_SIZE tx_size) { 416 const int max_blocks_high = max_block_high(xd, plane_bsize, plane) 417 << MI_SIZE_LOG2; 418 return ALIGN_POWER_OF_TWO(max_blocks_high, tx_size_high_log2[tx_size]); 419 } 420 421 void cfl_store_block(MACROBLOCKD *const xd, BLOCK_SIZE bsize, TX_SIZE tx_size) { 422 CFL_CTX *const cfl = &xd->cfl; 423 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y]; 424 int row = 0; 425 int col = 0; 426 427 if (block_size_high[bsize] == 4 || block_size_wide[bsize] == 4) { 428 sub8x8_adjust_offset(cfl, xd->mi_row, xd->mi_col, &row, &col); 429 } 430 const int width = max_intra_block_width(xd, bsize, AOM_PLANE_Y, tx_size); 431 const int height = max_intra_block_height(xd, bsize, AOM_PLANE_Y, tx_size); 432 tx_size = get_tx_size(width, height); 433 cfl_store(cfl, pd->dst.buf, pd->dst.stride, row, col, tx_size, 434 is_cur_buf_hbd(xd)); 435 }